River Extraction Method of Remote Sensing Image Based on Edge Feature Fusion

The extraction of rivers from remote sensing images is crucial for urban planning and water resource utilization. To address the low accuracy of traditional methods for extracting rivers from remote sensing images in complex scenes, a novel method based on edge feature fusion is proposed. As the input item of Learning Vector Quantization (LVQ), the preprocessed the remote sensing image is fed into four traditional edge detection algorithms for processing, and a finer edge image is obtained. Furthermore, VGG16 and ResNet50 models are utilized as feature extractors for transfer learning. The scale feature output layer and the semantic feature output layer are introduced to construct a Two-Branch Fusion (TBF) model. The scale feature output layer of the model makes full use of the rich convolution features of VGG16, and optimizes the training steps via collaborative loss and deep supervision. The semantic feature output layer extracts the residual maps matched by the ResNet50’s different scale convolution layers, and employs edge features to refine the segmentation boundary, thereby sharpening predictions at the edges. Moreover, the TBF model effectively integrates multi-scale edge information, thus improving the prediction accuracy of the model. The proposed model is tested with SVM, FCN, and UNet on the remote sensing river image dataset. The results show that the proposed model has greater PA (Pixel Accuracy), Pre (Precision), and mIoU (mean Intersection over Union) than other models. The test results indicate that the proposed model’s river segmentation effect is more refined. When applying this method to the Inria Aerial Image Labeling Dataset and Pascal VOC2012 Dataset, the model’s generalization ability is validated, and its evaluation index and segmentation effect are superior to those of other models.